Pointless for server to stay behind load balancer?

[Johan Jian An Sim]

Hi,

I am looking at the gRPC Go which logs the events/requests to net/trace. When I look at that, I get a hint that client are tied to a server. Want to see if I am wrong about this.

So if I have 3 servers ready (behind Kubernetes service/load balancer) and only 1 client is going to connect. Is the client's request always going to 1 server (the server that it establish connection initially)?

[Qi Zhao]

I have no idea how your balancer work. But by now all the rpcs are sent to whatever target you call grpc.Dial(...).

--
You received this message because you are subscribed to the Google Groups "grpc.io" group.
To unsubscribe from this group and stop receiving emails from it, send an email to grpc-io+u...@googlegroups.com.
To post to this group, send email to grp...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/grpc-io/204a73f5-7081-4d3c-9f8b-3b7fd6ca00c2%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--

Thanks,

-Qi

[Eric Anderson]

Each RPC is a different HTTP request, but generally on the same TCP connection (via HTTP/2 multiplexing). So if you are using an HTTP load balancer then requests would use your various servers. If you are using a TCP load balancer then all requests from a client would go to the same server, until that connection was broken.

Note that streaming RPC are each a single HTTP request; all request messages in the stream go to the same server.

On Tue, Oct 6, 2015 at 9:54 PM, Johan Jian An Sim <joha...@gmail.com> wrote:

--

[Johan Sim]

Thanks, guys. I will do some study on that.

[idan...@gmail.com]

Hi,

Found this thread while searching for an answer to the same issue...

I am doing the same implementation, having a client that is directed to a kubernetes service (DNS discovery) and I have noticed that all calls are going to the same server.

This is a single client with multiple servers behind the "service" identity.

Is there any way to bypass is so that each call will actually choose a random server in the service pool? 

[Kun Zhang]

If you are using Java, a proposed new LoadBalancer API should be able to solve your problem:

https://github.com/grpc/grpc-java/issues/1600

It will allow LoadBalancer to create multiple connections for the same address.

[Eric Anderson]

The tracking issue is https://github.com/grpc/grpc/issues/7957 . The first resolution (#1) would be to add a timer to the server to send a GOAWAY after the connection is a certain age, allowing the proxy to choose another backend.

I have an open question in that issue whether we also need an option to make multiple connections and RR on them hoping they go to different backends (#2). If you need this, I'd suggest commenting on the issue.